Unsupervised Language Model Weight Selection via Best Path Change Rate
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges in selecting an optimal language model weight (LMW) due to the need for similar development and test data, which is often difficult to collect and requires calculating word error rates (WER) for each LMW, making the process inefficient and unreliable.
Innovation Solution
An unsupervised method based on best path change rate (BPCR) is introduced, where test audio is decoded into a lattice using multiple LMWs, and the LMW with the lowest BPCR is selected as optimal, eliminating the need for similar development and test data and reducing the computational burden of WER calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional WER-based LMW selection method is used, then reliable LMW selection can be achieved, but it requires large amounts of similar development and test data which is difficult to collect
Solution Approach 1:
The system uses the test data itself to select the optimal LMW by analyzing lattice structures and best path changes, eliminating the need for separate development data. The test data serves dual purposes: both evaluation and optimization, allowing reliable LMW selection without requiring additional similar development datasets.
Solution Approach 2:
Instead of using WER (which requires reference transcriptions) to select LMW, the patent inverts the approach by using best path change rate analysis on the lattice structure itself. This inversion allows LMW selection without needing reference truths, thereby eliminating the requirement for separate development data while maintaining selection reliability.
2Measurement precision
If traditional WER-based LMW selection method is used, then accurate LMW selection can be achieved, but it requires calculating WER for every LMW which increases computational burden
Solution Approach 1:
The patent extracts the essential information needed for LMW selection directly from the lattice structure (best path changes) without performing full WER calculations. By taking out only the necessary lattice analysis components rather than computing complete WER metrics, the method maintains selection accuracy while significantly reducing computational power requirements.
Solution Approach 2:
Instead of performing complete WER calculations for every LMW candidate, the patent applies partial action by analyzing only the best path changes in the lattice. This partial analysis provides sufficient information for accurate LMW selection without the excessive computational cost of full WER computation, achieving the right balance between precision and efficiency.
3Reliability
If human-transcribed development data is used for LMW selection, then reliable selection can be achieved, but it increases time and resource requirements
Solution Approach 1:
The test data performs dual duty by serving both as the evaluation dataset and the optimization dataset for LMW selection. This self-service approach eliminates the need for separate human-transcribed development data, thereby maintaining selection reliability while significantly reducing the time and resources required for data preparation and processing.
Data Source
AI summary
A method for selecting an optimal language model weight (LMW) used to perform automatic speech recognition, including decoding test audio into a lattice using a language model; analyzing the lattice using a first LMW of a plurality of LMWs to determine a first plurality of best paths; analyzing the lattice using a second LMW of the plurality of LMWs to determine a second plurality of best paths; determining a first best path change rate (BCPR) corresponding to the first LMW based on a number of best path changes between the first plurality of best paths and the second plurality of best paths; and determining the first LMW to be the optimal LMW based on the first BCPR being a lowest BCPR from among a plurality of BCPRs corresponding to the plurality of LMWs.